Pareto Severity Maximum Likelihood

Introduction

The Pareto Type I distribution is a useful severity model for losses whose tail decreases as a power of the loss amount. In this tutorial, the threshold is known and only the shape parameter must be estimated from a complete sample. The essential calculation is simple once the likelihood is written correctly: the estimate is the sample size divided by the sum of the logarithms of the loss-to-threshold ratios.

“Complete” means that each sampled loss is observed at its actual value, without censoring or truncation. A deductible or reporting rule can change the likelihood; the formula below should not be applied to incomplete records as though they were complete. The focus here is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Pareto Type I Model

Definition: A loss $X$ has a Pareto Type I distribution with known threshold $\theta>0$ and shape parameter $\alpha>0$ if its support is $x\geq\theta$, its distribution function is $F_X(x)=1-(\theta/x)^\alpha$ for $x\geq\theta$, and its density is $f_X(x)=\alpha\theta^\alpha x^{-(\alpha+1)}$ for $x\geq\theta$. For $x<\theta$, $F_X(x)=0$ and $f_X(x)=0$.

Equivalently, for $x\geq\theta$, the survival function is $S_X(x)=(\theta/x)^\alpha$. The threshold is the minimum possible loss in the model, while the shape parameter controls how quickly the tail probability declines. A smaller $\alpha$ gives a heavier tail. In this tutorial $\theta$ is fixed by the model specification; it is not estimated from the sample.

Suppose $X_1,\ldots,X_n$ are independent and identically distributed complete observations from this model. The observations must satisfy $x_i\geq\theta$. Define $T=\sum_{i=1}^{n}\ln(x_i/\theta)$, where $\ln$ denotes the natural logarithm. If at least one observation is strictly above $\theta$, then $T>0$.

Likelihood and Score Equation

Independence means that the joint density is the product of the individual densities. For observations all within the support, the likelihood as a function of $\alpha$ is

L(\alpha)=\prod_{i=1}^{n}\alpha\theta^\alpha x_i^{-(\alpha+1)} =\alpha^n\theta^{n\alpha}\prod_{i=1}^{n}x_i^{-(\alpha+1)},\qquad \alpha>0.

The product of the $x_i$ terms does not depend on $\alpha$ when separated appropriately. Taking logarithms converts the product into a sum and gives the log-likelihood

\ell(\alpha)=\ln L(\alpha) =n\ln\alpha-\alpha T-\sum_{i=1}^{n}\ln x_i.

The final sum is constant with respect to $\alpha$. Differentiating the log-likelihood and setting the score equal to zero gives

\frac{d\ell}{d\alpha}=\frac{n}{\alpha}-T=0 \quad\Longrightarrow\quad \widehat{\alpha}=\frac{n}{T} =\frac{n}{\displaystyle\sum_{i=1}^{n}\ln(x_i/\theta)}.

Theorem: For a complete independent sample from a Pareto Type I model with known threshold $\theta$, if all observations satisfy $x_i\geq\theta$ and at least one is strictly greater than $\theta$, the maximum likelihood estimate of the shape parameter is $\widehat{\alpha}=n/\sum_i\ln(x_i/\theta)$. It is the unique maximum because $\ell''(\alpha)=-n/\alpha^2<0$ for every $\alpha>0$.

Why the Estimate Has This Form

There is also a helpful transformation that explains the score equation. Let $Z=\ln(X/\theta)$. For $z\geq0$,

P(Z>z)=P(X>\theta e^z) =\left(\frac{\theta}{\theta e^z}\right)^\alpha =e^{-\alpha z}.

Thus $Z$ has an exponential distribution with rate $\alpha$. For an independent sample, the transformed observations $z_i=\ln(x_i/\theta)$ have joint likelihood proportional to $\alpha^n e^{-\alpha\sum z_i}$. Its log-likelihood has score $n/\alpha-\sum z_i$, exactly the score obtained above. The estimate is therefore the exponential rate estimate applied to the transformed losses.

This argument also clarifies why the estimate uses logarithms of ratios, not logarithms of losses alone: the transformation must compare each loss with the model’s threshold. The threshold is what makes $x_i/\theta$ dimensionless.

Worked Applications

Worked Example: Estimating Shape From Three Complete Losses

A hypothetical portfolio has complete losses of $1,500, $3,000, and $6,000, modeled as independent Pareto Type I losses with known threshold $\theta=\$1{,}000$ and unknown shape $\alpha$. Estimate $\alpha$ to four decimal places. No interest or mortality assumption is involved; this is a severity-only model.

Method. Use the complete-data Pareto Type I score estimate. First check support: $1,500, $3,000, and $6,000 are all at least the $1,000 threshold.

The loss-to-threshold ratios are $1{,}500/1{,}000=1.5$, $3{,}000/1{,}000=3$, and $6{,}000/1{,}000=6$. Therefore,

T=\ln(1.5)+\ln(3)+\ln(6) =\ln(1.5\cdot3\cdot6) =\ln(27) \approx3.295836866.

There are $n=3$ losses, so

\widehat{\alpha}=\frac{3}{3.295836866} \approx0.9102392.

Rounded only at the final step, the estimate is $\widehat{\alpha}=0.9102$. As a check, $3/\widehat{\alpha}$ is approximately $3.2958369$, the calculated value of $T$. A common mistake is to report 0.9104; that does not result from dividing 3 by $\ln(27)$.

Worked Example: Working With Exact Multiples of the Threshold

A hypothetical equipment insurer observes complete losses of $4,000, $8,000, and $16,000. Assume independent Pareto Type I losses with known threshold $\theta=\$2{,}000$ and unknown shape $\alpha$. Find the maximum likelihood estimate to four decimal places. No interest or mortality assumption is involved.

Method. Divide every observed loss by the known threshold before taking logarithms. The ratios are $4{,}000/2{,}000=2$, $8{,}000/2{,}000=4$, and $16{,}000/2{,}000=8$. All three observations are within the support.

The log-ratio sum can be computed by adding the logarithms or by multiplying the ratios:

T=\ln 2+\ln 4+\ln 8 =\ln(2\cdot4\cdot8) =\ln(64) \approx4.158883083.

With $n=3$,

\widehat{\alpha} =\frac{3}{4.158883083} \approx0.7213475 \approx0.7213.

To check the score, $n/\widehat{\alpha}$ is approximately $4.1588831$, matching $T$. The second derivative is $-3/\widehat{\alpha}^2<0$, so this stationary point is a maximum. The common mistake avoided here is treating the threshold as though it were an additional unknown parameter; it is specified as $2,000 and is used only to form the ratios.

Worked Example: Checking Support Before Estimating

A hypothetical coverage produces complete recorded losses of $1,000, $1,500, $3,000, and $6,000. Assume a Pareto Type I model with known threshold $\theta=\$500$ and unknown shape $\alpha$. Estimate $\alpha$ to four decimal places. No interest or mortality assumption is involved.

Method. Check the support, then apply the score estimate. The observations are all at least $500. Their ratios to the threshold are $2,3,6,$ and $12$, so

T=\ln 2+\ln 3+\ln 6+\ln 12 =\ln(2\cdot3\cdot6\cdot12) =\ln(432) \approx6.068425588.

There are four observations:

\widehat{\alpha} =\frac{4}{6.068425588} \approx0.6591496 \approx0.6591.

For a numerical score check, $4/0.6591496\approx6.06843$, agreeing with $T$ to the shown precision. The common mistake avoided is taking logarithms of the dollar amounts without dividing by $500 first. That would not give the Pareto score equation and would also make the calculation depend on the units used for the losses.

Conditions and Common Exam Traps

The formula is a maximum likelihood estimate only under the stated model and data conditions. It uses a known threshold, independent observations, and complete losses from the Pareto Type I distribution. If a loss is reported only because it exceeded a deductible, or if a loss is censored at a policy limit, its contribution to the likelihood differs from an observed complete density. Treating such a record as complete would change the likelihood and generally change the estimate.

Do not estimate the threshold from this score equation. It assumes $\theta$ is known. A different problem with an unknown threshold requires a likelihood that treats the threshold as an unknown parameter and accounts for its support constraints.
Check every observation against the threshold. If any $x_i<\theta$, the specified model assigns that observation density zero, so the sample is incompatible with the assumed model. Do not insert the invalid loss into the formula.
Use natural logarithms of ratios. Each term is $\ln(x_i/\theta)$. The order matters: $x_i\geq\theta$, so the ratios are at least one and the terms are nonnegative.
Do not confuse the score equation with the log-likelihood. The score is $n/\alpha-T$; setting it to zero gives $n/\alpha=T$, not $\alpha=nT$.
Recognize the degenerate boundary case. If every observation equals $\theta$, then $T=0$ and there is no finite estimate from the formula. The likelihood is $(\alpha/\theta)^n$, which increases without bound as $\alpha$ increases. A finite maximum does not exist in this case.

For any valid sample with $T>0$, the score changes from positive to negative at $n/T$: if $\alpha<n/T$, then $n/\alpha>T$, while if $\alpha>n/T$, then $n/\alpha<T$. Together with the strictly negative second derivative, this confirms that the estimate is the unique maximum rather than merely a solution to the score equation.

Key takeaway: For complete Pareto Type I losses with known threshold $\theta$, calculate the natural-log sum of the loss-to-threshold ratios and divide the sample size by that sum. First verify that every loss is at least $\theta$ and that the sum is positive.

Check Your Understanding

For each hypothetical sample, assume independent complete Pareto Type I losses with the stated known threshold and unknown shape. Give estimates to four decimal places unless the data are incompatible with the model.

The known threshold is $1,000 and the observed losses are $2,000, $4,000, and $8,000. Find the shape estimate. (Answer: $\widehat{\alpha}=0.7213$)
The known threshold is $500 and the losses are $750, $1,000, and $2,000. Find the shape estimate. (Answer: $\widehat{\alpha}=1.2071$)
The known threshold is $2,500 and the losses are $5,000, $7,500, $10,000, and $20,000. Find the shape estimate. (Answer: $\widehat{\alpha}=0.7608$)
The known threshold is $100 and the losses are $200, $300, $600, and $1,200. Find the shape estimate, showing the log-ratio sum. (Answer: $T=\ln(432)\approx6.0684$, $\widehat{\alpha}=0.6591$)
The known threshold is $1,000 and the recorded complete losses are $800, $1,500, and $3,000. Can the stated Pareto Type I model produce these data, and is there a valid finite estimate from the formula? (Answer: No; the $800 loss is below the threshold, so the sample is incompatible with the model and the formula does not give a valid estimate.)

What You'll Learn